From: Jesse Zhang <Jesse.Zhang@amd.com>
To: <amd-gfx@lists.freedesktop.org>
Cc: <Alexander.Deucher@amd.com>,
Christian Koenig <christian.koenig@amd.com>,
Jesse.zhang <Jesse.zhang@amd.com>,
"David (Ming Qiang) Wu" <David.Wu3@amd.com>,
Alex Deucher <alexander.deucher@amd.com>,
Jesse Zhang <Jesse.Zhang@amd.com>
Subject: [PATCH v2 06/11] drm/amdgpu: add AMDGPU_INFO_USERQ_DOORBELL ioctl
Date: Mon, 27 Apr 2026 16:34:32 +0800 [thread overview]
Message-ID: <20260427083543.1328533-6-Jesse.Zhang@amd.com> (raw)
In-Reply-To: <20260427083543.1328533-1-Jesse.Zhang@amd.com>
From: "Jesse.zhang" <Jesse.zhang@amd.com>
Per-IP doorbell layout query. Input: info->query_hw_ip.type selects
the IP (AMDGPU_HW_IP_GFX / COMPUTE / DMA / VCN_ENC). Output: the
doorbell-BAR dword range usable by usermode queues for that IP, the
aggregated doorbell offset (when MES owns one), and — for
AMDGPU_HW_IP_DMA only — a per-fpriv GEM handle plus byte size for the
kernel-owned BO that backs the routable SDMA doorbell window.
For SDMA usermode queues the kernel hands out doorbell slots from the
per-device pool (the only BAR range whose writes are routed to the
SDMA back-end) instead of accepting user-allocated doorbell BOs.
Userspace mmap()s the returned handle through the standard
AMDGPU_GEM_OP_MMAP / mmap() flow to obtain a CPU pointer to the
window; each created SDMA UMQ's qword-slot offset inside that mapping
is reported in drm_amdgpu_userq_out.sdma_doorbell_offset_bytes.
V2: Picks up David Wu's [PATCH 11/14] AMDGPU_INFO_DOORBELL design (struct
shape + per-IP dispatch) and extends it with the SDMA-only routable-BO
handle/size fields.
Cc: David (Ming Qiang) Wu <David.Wu3@amd.com>
Suggested-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: David (Ming Qiang) Wu <David.Wu3@amd.com>
Signed-off-by: Jesse Zhang <Jesse.Zhang@amd.com>
---
drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c | 35 ++++++++++++++++++++
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c | 13 ++++----
include/uapi/drm/amdgpu_drm.h | 40 +++++++++++++++++++++--
3 files changed, 78 insertions(+), 10 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c
index d88e4994c8c1..32adcb32a507 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c
@@ -1425,6 +1425,41 @@ int amdgpu_info_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
return -EINVAL;
}
}
+ case AMDGPU_INFO_USERQ_DOORBELL: {
+ struct drm_amdgpu_info_userq_doorbell db_info = {};
+ u32 agdb = adev->enable_mes ?
+ adev->mes.aggregated_doorbells[AMDGPU_MES_PRIORITY_LEVEL_NORMAL] : 0;
+ int r;
+
+ switch (info->query_hw_ip.type) {
+ case AMDGPU_HW_IP_DMA:
+ if (!adev->sdma.userq_db_obj)
+ return -ENODEV;
+ db_info.index_start =
+ adev->doorbell_index.sdma_engine[0] << 1;
+ db_info.index_end = db_info.index_start +
+ adev->doorbell_index.sdma_doorbell_range *
+ adev->sdma.num_instances - 1;
+ if (agdb) {
+ db_info.agdb_enable = 1;
+ db_info.agdb_offset = agdb;
+ }
+ r = amdgpu_sdma_userq_doorbell_create_handle(adev, filp,
+ &db_info.doorbell_bo_handle,
+ &db_info.doorbell_bo_size_bytes);
+ if (r)
+ return r;
+ break;
+ case AMDGPU_HW_IP_VCN_ENC:
+ case AMDGPU_HW_IP_GFX:
+ case AMDGPU_HW_IP_COMPUTE:
+ default:
+ return -EINVAL;
+ }
+ return copy_to_user(out, &db_info,
+ min((size_t)size, sizeof(db_info)))
+ ? -EFAULT : 0;
+ }
default:
DRM_DEBUG_KMS("Invalid request %d\n", info->query);
return -EINVAL;
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
index cea0f9cb59d0..a0c90cc0cba5 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
@@ -804,10 +804,10 @@ amdgpu_userq_create(struct drm_file *filp, union drm_amdgpu_userq *args)
* doorbell index with one allocated from the per-device
* sdma.userq_db_obj BO that sits inside the routable window.
* Userspace fetches a GEM handle for that BO via
- * AMDGPU_INFO_SDMA_USERQ_DOORBELL and mmap()s it through the
- * standard GEM_MMAP path; sdma_doorbell_offset_bytes (returned in
- * args->out) tells userspace where inside that mapping its slot
- * lives.
+ * AMDGPU_INFO_USERQ_DOORBELL with AMDGPU_HW_IP_DMA and mmap()s it
+ * through the standard GEM_MMAP path; sdma_doorbell_offset_bytes
+ * (returned in args->out) tells userspace where inside that
+ * mapping its slot lives.
*/
if (queue->queue_type == AMDGPU_HW_IP_DMA &&
adev->sdma.userq_db_obj) {
@@ -891,9 +891,8 @@ amdgpu_userq_create(struct drm_file *filp, union drm_amdgpu_userq *args)
if (queue->sdma_userq_db_slot >= 0) {
/*
* Tell userspace where inside its mmap of the SDMA UMQ
- * doorbell BO (handle returned by
- * AMDGPU_INFO_SDMA_USERQ_DOORBELL) the assigned qword slot
- * lives.
+ * doorbell BO (handle returned by AMDGPU_INFO_USERQ_DOORBELL with
+ * AMDGPU_HW_IP_DMA) the assigned qword slot lives.
*/
args->out.sdma_doorbell_offset_bytes =
(u64)queue->sdma_userq_db_slot * sizeof(u64);
diff --git a/include/uapi/drm/amdgpu_drm.h b/include/uapi/drm/amdgpu_drm.h
index 79e8bbda046b..945fa3e95b3b 100644
--- a/include/uapi/drm/amdgpu_drm.h
+++ b/include/uapi/drm/amdgpu_drm.h
@@ -423,9 +423,10 @@ struct drm_amdgpu_userq_out {
__u32 _pad;
/**
* For SDMA usermode queues whose doorbell was assigned by the
- * kernel from the per-device pool (see AMDGPU_INFO_SDMA_USERQ_DOORBELL),
- * this field carries the byte offset of the assigned slot inside
- * the routable doorbell window so userspace can write there.
+ * kernel from the per-device pool (see AMDGPU_INFO_USERQ_DOORBELL with
+ * AMDGPU_HW_IP_DMA), this field carries the byte offset of the
+ * assigned slot inside the routable doorbell window so userspace
+ * can write there.
* 0 means the kernel did not override the user's doorbell.
*/
__u64 sdma_doorbell_offset_bytes;
@@ -1280,6 +1281,39 @@ struct drm_amdgpu_cs_chunk_cp_gfx_shadow {
#define AMDGPU_INFO_GPUVM_FAULT 0x23
/* query FW object size and alignment */
#define AMDGPU_INFO_UQ_FW_AREAS 0x24
+/*
+ * Per-IP doorbell layout query. Input: info->query_hw_ip.type selects
+ * the IP (AMDGPU_HW_IP_GFX / COMPUTE / DMA / VCN_ENC). Output:
+ * doorbell-BAR dword range usable by usermode queues for that IP, the
+ * aggregated doorbell offset (when MES/UMSCH owns one for the IP), and
+ * — for AMDGPU_HW_IP_DMA only — a per-fpriv GEM handle for a
+ * kernel-owned BO that backs the routable SDMA doorbell window so
+ * userspace can mmap() its assigned slot.
+ */
+#define AMDGPU_INFO_USERQ_DOORBELL 0x25
+
+struct drm_amdgpu_info_userq_doorbell {
+ /* BAR dword index of the start of the per-IP doorbell range. */
+ __u32 index_start;
+ /* BAR dword index of the last doorbell in the range (inclusive). */
+ __u32 index_end;
+ /* 1 if an aggregated doorbell exists for this IP. */
+ __u32 agdb_enable;
+ /* If agdb_enable, BAR dword index of the aggregated doorbell. */
+ __u32 agdb_offset;
+ /*
+ * AMDGPU_HW_IP_DMA only: per-fpriv GEM handle for the kernel-owned
+ * BO backing the routable SDMA doorbell window. Userspace mmap()s
+ * it through the standard AMDGPU_GEM_OP_MMAP / mmap() flow to get
+ * a CPU pointer; each created SDMA usermode queue's qword-slot
+ * offset inside that mapping is reported in
+ * drm_amdgpu_userq_out.sdma_doorbell_offset_bytes. 0 for IPs that
+ * do not need a kernel-managed doorbell BO.
+ */
+ __u32 doorbell_bo_handle;
+ /* Byte size of the BO; 0 when doorbell_bo_handle is 0. */
+ __u32 doorbell_bo_size_bytes;
+};
#define AMDGPU_INFO_MMR_SE_INDEX_SHIFT 0
#define AMDGPU_INFO_MMR_SE_INDEX_MASK 0xff
--
2.49.0
next prev parent reply other threads:[~2026-04-27 8:36 UTC|newest]
Thread overview: 14+ messages / expand[flat|nested] mbox.gz Atom feed top
2026-04-27 8:34 [PATCH v2 01/11] drm/amdgpu/sdma: add SDMA usermode-queue doorbell pool infra Jesse Zhang
2026-04-27 8:34 ` [PATCH v2 02/11] drm/amdgpu/userq: route SDMA UMQ doorbells through the kernel pool Jesse Zhang
2026-04-27 8:34 ` [PATCH v2 03/11] drm/amdgpu/gem: only enforce amdgpu_bo access checks on amdgpu_bo objects Jesse Zhang
2026-04-27 8:39 ` Christian König
2026-04-27 8:34 ` [PATCH v2 04/11] drm/amdgpu/sdma7: register SDMA UMQ doorbell pool Jesse Zhang
2026-04-27 8:34 ` [PATCH v2 05/11] drm/amdgpu/sdma6: " Jesse Zhang
2026-04-27 8:34 ` Jesse Zhang [this message]
2026-04-27 8:34 ` [PATCH v2 07/11] drm/amdgpu/mes: add NOTIFY_WORK_ON_UNMAPPED_QUEUE op + ADD_QUEUE fields Jesse Zhang
2026-04-27 8:34 ` [PATCH v2 08/11] drm/amdgpu/mes11: plumb unmap_flag_addr + NOTIFY_WORK_ON_UNMAPPED_QUEUE Jesse Zhang
2026-04-27 8:34 ` [PATCH v2 09/11] drm/amdgpu/mes12: plumb is_user_mode_submission, unmap_flag_addr, NOTIFY Jesse Zhang
2026-04-27 8:34 ` [PATCH v2 10/11] drm/amdgpu/mes_userqueue: mark SDMA UMQs as user-mode submission Jesse Zhang
2026-04-27 8:34 ` [PATCH v2 11/11] drm/amdgpu/userq_fence: wake gangs-out SDMA UMQs via NOTIFY Jesse Zhang
2026-04-27 8:42 ` [PATCH v2 01/11] drm/amdgpu/sdma: add SDMA usermode-queue doorbell pool infra Christian König
2026-04-28 9:39 ` Zhang, Jesse(Jie)
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=20260427083543.1328533-6-Jesse.Zhang@amd.com \
--to=jesse.zhang@amd.com \
--cc=Alexander.Deucher@amd.com \
--cc=David.Wu3@amd.com \
--cc=amd-gfx@lists.freedesktop.org \
--cc=christian.koenig@amd.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox